Tarot-Draw
Laplace Research ↗A family of foundation models for calibrated forecasting.
Selected work
Research
Independent Research
Reward Variation Without Demonstrated Token Savings: An Audited Exploratory Study
Examines how a token-cost penalty activates additional policy-gradient updates in a small expression-repair model without demonstrating lower final token use, with mathematical derivations, audited measurements, and explicit limitations.
Climate Information and Wheat Futures: A Filtered-Information Theory of Forecast Value, Equilibrium Response, and Trading Feasibility
Develops a mathematical theory of climate information in wheat futures, with proofs for physical forecast value, equilibrium price sensitivity, the nonidentification of trading alpha, and cost-aware abstention. Includes full derivations and a point-in-time testing protocol; no completed backtest or profitability claim.
An Investigation Regarding Pseudorandom Number Generation with Xor Shifting
Explores xorshift pseudorandom number generation through a mathematical model and a 20,000-number sample. This was my Internal Assessment for IB Mathematics HL 2.
Laplace Research
Tarot-Draw Technical Report
Documents how Tarot-Draw was built and evaluated, including training variance, calibration and resolution, compute cost, and a dated record of 51 corrections.
Manufactured Skill in Prediction-Market Forecasting
Shows how a common preprocessing step can make prediction-market forecasts appear more skillful than they are, using a benchmark built from two venues.
Coherent, Not Merely Accurate
Uses no-arbitrage constraints across related prediction-market contracts to reveal forecast quality that accuracy metrics can miss.
Crypto Research
Dippi: Architecture, Trust Boundaries, and Reproducible Measurement of a Testnet Spot Exchange
Examines Dippi's matching engine, ledger, custody, and recovery, with repeated CPU benchmarks, a locally tested repair for a cross-market balance regression, and explicit measurement limits.